next up previous
Next: Conclusion Up: Results Previous: Area-Oriented Comparison with UCLOCK

Exploring Varying Code Lengths

This section briefly shows the effect on the synthesis results for a single design when varying code length in MINIMALIST.

Because a simple cost metric often fails to capture an application's cost completely, MINIMALIST better assists the designer in finding the point which best fits the application, by providing the opportunity to explore trade-offs. For example, Table 3 shows an interesting trade-off involving code length, arising from two competing tendencies. Output logic tends to improve, while next-state logic tends to grow, with increasing code length.

For these runs, a single design (scsi-tsend-csm) is synthesized using the performance-oriented script, but targetting the fed-back output machine implementation style. Code length is varied from 3 (the minimum needed to ensure a critical race-free encoding) to 6 (one less than the code length resulting when CHASM uses its exact, rather than its fixed-length, mode).

As the results show, the output logic complexity decreases as the code length increases, in exchange for a more expensive next-state implementation. This reflects the fact that CHASM'S input encoding model is exact for outputs, but is only approximate for next-state. Specifically, the fixed-length constraint satisfaction method favors neither output nor next-state constraints [8]. Thus, longer codes tend to satisfy a greater number of both kinds of constraints. So, output logic complexity decreases, because the corresponding constraints precisely model logic optimality. However, the next-state, whose constraints are less accurate, experiences an increase in logic complexity.

Without the ability to explore such trade-offs, a designer is forced to choose whatever single point in the solution space the synthesis path chooses. For example, suppose a synthesis path always chose to minimize output literal count. Given the results in Table 3, this would force the designer to accept an \( 8\% \) decrease in output logic complexity, in exchange for a \( 71\% \) increase in next-state logic complexity, which might be intolerable. MINIMALIST's ability to explore such trade-offs is unique among burst-mode synthesis toolkits.


 
Table 3: Effect of varying code length on synthesis results for a single design


design in/state/out codelen outprods outlits nsprods nslits totprods totlits
scsi-tsend-csm 5/10/4 3 14 37 9 24 23 61
`` `` 4 14 36 7 26 21 62
`` `` 5 13 35 10 34 23 69
`` `` 6 13 34 11 41 24 75





next up previous
Next: Conclusion Up: Results Previous: Area-Oriented Comparison with UCLOCK
Steven Nowick
1999-07-28